EverOS is EverMind’s Apache-2.0 local-first memory runtime for agents: conversations, files, and agent trajectories are stored as editable Markdown files that remain the source of truth, with SQLite and LanceDB indexes built beside them for retrieval, and a managed cloud offered above the same stack.
Its bet is that memory should be files you can open and edit, with the indexes as rebuildable views, and that user memory and agent memory are two different products worth separate tracks.
What it is #
A Python library and local server (pip install everos, Python 3.12+), startable keyless for the demo and with a single OpenRouter key for the full loop; the local stack is Markdown plus SQLite plus LanceDB with no external database. Memory is organized on two first-class tracks: user episodes and profiles, and agent cases and skills distilled from trajectories, retrieved through ids scoped by user, agent, app, project, and session. A reflection pass merges episode clusters and refines profiles and skills between sessions, and a knowledge-wiki surface exposes editable, source-backed Markdown pages. Ingestion is multimodal through the vendor’s mRAG layer (PDFs, images, documents, URLs). Distribution leans on harness plugins: DeepSeek Harness, Hermes, and OpenClaw carry official EverOS plugins, Dify is supported, and EverOS is built into the vendor’s own Raven agent. EverMind also sells EverOS Cloud, a managed deployment of the same engine, and publishes a research line (EverCore, with the EverMemOS and HyperMem papers) behind the architecture.
Status #
Young, fast-moving, and vendor-led, with the section’s familiar thin public footprint. 13,444 stars and 941 forks as of 2026-10-28, created 2025-10-28, default branch pushed 2026-11-19 (GitHub API, re-checked 2026-11-19).
Latest release v1.7.0 (2026-11-13) with PyPI in sync at 1.7.0, a 1.x line that reached 1.7 within a year of creation: the pending 1.6.0 contract-and-refactor line shipped (the everalgo pinned algorithm packages, the MemoryService, KnowledgeService, and OpsService composition-root refactor, and the breaking settings changes recorded on 2026-10-31) alongside the Cloud-aligned memory deletion endpoints (deletion by user_id/agent_id, by session, or one owner’s session, answering with the matched filters and a count, refusing with SESSION_BUSY while a session is still being memorized, and everos-mcp’s forget_session now working against a self-hosted server), recovery improvements when writes, background tasks, and file updates overlap, safeguards for large search workloads, and unused vectors kept out of query results; v1.5.0 (October 21) had added async ingestion (an /add with async_mode answers in milliseconds and extracts on a durable background queue, matching the cloud API’s contract), configurable embedding dimensions, and startup validation that refuses an index the configured model cannot have produced, and v1.5.1 (October 22) raised the default query-embedding concurrency from 4 to 16, lifting sustained hybrid search from about 9 to a held 10 requests a second.
The tracked 1.7.0 line shipped on November 13: the pending line’s breaking everalgo 2026-11 contract with per-task LLM and profile building from episodes (every profile item cites the conversation it comes from) plus the online purge API (POST /api/v2/memory/purge with a confirm step and a job-status endpoint, clearing a whole memory root, one app, or one project while the server keeps serving).
The pending line grew again (November 16, breaking): one EpisodeExtracted record per memcell with episode summaries built offline, still unreleased on GitHub and PyPI (both at v1.7.0/1.7.0).
The pending line grew again (November 18 to 19, breaking): search rebuilt as three-lane hybrid ranking with agentic facts capped, then body-vector agentic search with bounded recall pools; GitHub and PyPI re-verified at v1.7.0/1.7.0.
The vendor’s site claims SOTA numbers (LoCoMo 93.05, LongMemEval 83.00, HaluMem 93.04) from its own reproducible benchmark harness, and the system does not appear among the published commercial leaders of the independent Agent Memory Leaderboard’s first cycle (August 2026).
I found no Hacker News thread through 2026-10-09 searches, so the 13k stars rest on the vendor’s channels and plugin ecosystems, the same stars-ahead-of-discussion pattern this section has flagged on Supermemory, Cabinet, and memU.
Strengths #
- The markdown-canonical local stack works as described: files are the store, the SQLite and LanceDB indexes are rebuildable views, and editing a file syncs back through the watcher.
- Separate user and agent tracks (episodes and profiles versus cases and skills) match how personal assistants and coding agents actually differ.
- The official plugin line covers OpenClaw, Hermes, and DeepSeek Harness from one core, the same distribution bet MemOS makes with a different storage model.
- A visible release train with PyPI in sync, against a category where several members ship through separate channels.
Cautions #
- Every benchmark number is the vendor’s own run of its own harness, the self-published pattern this section flags on Mem0, MemOS, and Supermemory, and the one independent common-framework evaluation in the field lists no EverOS result among its published leaders.
- The README’s checkmark-and-cross competitor table and the site’s “SOTA” framing are marketing, not evaluation.
- The 1.x line is young, and the hosted cloud’s pricing is unpublished: the site’s pricing page returned 404 at fetch time on 2026-10-09, so the managed tier is a contact-sales path today.
- The public footprint is thin (no HN record), so failure reports and fixes skew toward the vendor’s Discord and Reddit channels.
- The quick start assumes an OpenRouter key for anything beyond the demo, a soft funnel toward one provider.
Pricing #
The open-source runtime is free, Apache-2.0, self-hosted on your own infrastructure. EverOS Cloud exists as a managed tier with no published price found as of 2026-10-09 (the pricing page returned 404 and the console sits behind sign-in), so it is an engagement, not a plan ladder.
Compared to #
- Memoryfields: both make Markdown canonical; Memoryfields is a portable format spec with minimal tooling, EverOS is a working runtime with indexes, reflection, and plugins attached.
- MemOS: the other plugin-carrying vendor memory OS; MemOS upgrades to Neo4j and Qdrant graphs, EverOS keeps everything in local files and indexes.
- memU: both bet on markdown the agent can read and edit; memU lets the agent curate the wiki, EverOS runs vendor-written extraction and reflection.
Bottom line #
Recommended for OpenClaw, Hermes, or DeepSeek Harness users who want a files-first memory runtime with user and agent tracks, and who will accept a young, vendor-led dependency. Not for teams needing published hosted pricing, independent benchmark evidence, or a proven multi-year release history.
Changes #
- 2026-10-09 - Created from the 2026-10-09 awesome-list entrant scan, with nine fetched sources and the self-published-benchmark-plus-thin-footprint combination recorded as the critical angle.
- 2026-10-16 - Recorded v1.4.2 (October 14), the durability patch for file-descriptor leaks, search queueing behind the embedding cascade, and multi-project index overwrites, with PyPI in sync at 1.4.2; refreshed stars to 13,444.
- 2026-10-19 - Corrected the PyPI reference line to 1.4.2, which the 2026-10-16 recording had already stated.
- 2026-10-24 - Recorded v1.5.0 (October 21, async ingestion with
async_modeon/addbacked by a durable queue, configurable embedding dimensions, startup index validation) and v1.5.1 (October 22, query-embedding concurrency default raised from 4 to 16), with PyPI in sync at 1.5.1. - 2026-10-27 - Rotation re-verification: every cited URL re-fetched live (the Agent Memory Leaderboard site, which returned 522 at the 2026-10-11 check, now fetches normally and its fetched surface still lists no EverOS result); no release after v1.5.1, PyPI in sync at 1.5.1, no facts changed.
- 2026-10-31 - Rotation re-verification: every cited URL re-fetched live (all nine reachable, no dead sources); recorded the breaking 1.6.0 landing on main (2026-10-30): the everalgo 2026-10 contract line with exactly pinned algorithm packages, and the target-architecture refactor (one composition root, MemoryService/KnowledgeService/OpsService facades, 14 import-linter contracts), breaking the cascade and OME environment variables into settings, requiring
[llm], moving prompt slots to operator YAML and service imports to the facades; still unreleased on GitHub and PyPI (both at v1.5.1/1.5.1) as of this run. - 2026-11-04 - The pending 1.6.0 line gained memory deletion endpoints aligned with the Cloud API (November 4), still unreleased on GitHub and PyPI (both at v1.5.1/1.5.1); the push date refreshed to 2026-11-04.
- 2026-11-08 - The pending line grew again (November 5-6): atomic session writes under contention, a root lock held before shared storage, cascade duplicate-index reporting with tombstone rechecks and claim-reset re-enqueues, reflection counting members that joined mid-run, a Windows markdown read-race retry, and the EverMemOS ACL 2026 paper cited in the docs; still unreleased on GitHub and PyPI (both at v1.5.1/1.5.1), the push date refreshed to 2026-11-07.
- 2026-11-11 - Recorded v1.6.0 (November 11, PyPI in sync at 1.6.0): the tracked pending line shipped (the pinned-algorithm contract, the service-facade refactor, the Cloud-aligned deletion endpoints with the SESSION_BUSY guard and self-hosted forget_session, the atomic-write and recovery work, and the search safeguards, with unused vectors kept out of results); the push date refreshed to 2026-11-11.
- 2026-11-13 - Volatile re-verification: the pending 1.7.0 line recorded (November 12-13, breaking: the everalgo 2026-11 contract with per-task LLM and episode-derived profiles, plus online purge of a memory root, app, or project); GitHub and PyPI still at v1.6.0/1.6.0, and the push date refreshed to 2026-11-13.
- 2026-11-14 - Recorded v1.7.0 (November 13, PyPI in sync at 1.7.0): the tracked pending line shipped (the breaking everalgo 2026-11 contract with per-task LLM, profile items citing their source conversation, and the online purge API with a confirm step and job-status endpoint).
- 2026-11-17 - Volatile re-verification: the pending line grew again (November 16, breaking): one EpisodeExtracted record per memcell with episode summaries built offline; GitHub and PyPI re-verified at v1.7.0/1.7.0, and the push date refreshed to 2026-11-16.
- 2026-11-19 - Volatile re-verification: the pending line grew again (November 18 to 19, breaking): search rebuilt as three-lane hybrid ranking with agentic facts capped, then body-vector agentic search with bounded recall pools; GitHub and PyPI re-verified at v1.7.0/1.7.0, and the push date refreshed to 2026-11-19.
See also #
- Memory Feature Matrix - the category comparison this note joins
- MemOS - the graph-upgrading rival for the same harness-plugin audience
- Memoryfields - the format-spec sibling of the markdown-canonical position
- memU - the agent-curated markdown wiki alternative
- File-based agent memory - the convention EverOS productizes with indexes and plugins
References #
https://github.com/EverMind-AI/EverOS - repository, 13,400 stars, forks, activity, description, as of 2026-10-11
https://raw.githubusercontent.com/EverMind-AI/EverOS/main/README.md - the markdown-plus-SQLite-plus-LanceDB stack, the two tracks, scoping ids, reflection, and the integration table
https://api.github.com/repos/EverMind-AI/EverOS/license - the Apache-2.0 license file, verified through the GitHub API
https://api.github.com/repos/EverMind-AI/EverOS/releases - the v1.4.x release train, as of 2026-10-09
https://pypi.org/pypi/everos/json - everos 1.4.2, in sync with the GitHub tag (re-verified 2026-10-19)
https://docs.evermind.ai - the documentation root: EverOS Cloud, open-source, and API reference surfaces
https://evermind.ai - the vendor site: SOTA claims, the EverCore research line, the product family, and the missing published pricing
https://hn.algolia.com/api/v1/search?query=everos&tags=story - the footprint scan finding no thread (2026-10-09), the thin-discussion record
https://agentmemoryleaderboard.ai/ - the independent academic-consortium leaderboard whose published first-cycle commercial leaders do not include this system (the site returned 522, origin unreachable, at fetch time on 2026-10-11)